> On September 11, 2026, Agibot (Zhiyuan Robotics) open-sourced GE-Act 2.0, a native world-action model pre-trained from random initialization on robot manipulation data. The team scaled joint training data from 300 hours to 30,000 hours, lifting the humanoid G1-OP's zero-shot success rate from 17.1% to 44.1%, and the number of tasks with non-zero scores from 39 to 76. 300 hours is 12.5 days; 30,000 hours is 3.4 years. This curve was repeatedly validated on language models, but robotics had lacked systematic evidence — until now.
Separating "World-Action Models" from VLAs
Standard vision-language-action models (VLAs) map camera frames and language instructions directly to motor commands — treating physical interaction as text continuation. Their weakness: they never explicitly model "what happens to the scene when I touch it."
A world-action model (WAM) fills in that gap: first predict how the scene should evolve, then decide how to act. The problem is that conventional video generation models are too slow and memory-hungry for real-time control loops.
GE-Act 2.0 splits the task into three stages:
- CoAE — a control-first vision tokenizer that compresses each 256×384 frame into 24 visual tokens, one-sixteenth of the common DINOv3 representation. It discards background detail while preserving geometry, boundaries, and motion. In a controlled test on 4,000 manipulation samples, it achieved 97.95% instruction-frame matching accuracy, the highest among five comparison representations.
- SVP — a one-step visual planner that generates future frames in a single forward pass instead of dozens of denoising iterations. Near frames drive immediate control; far frames keep the trajectory on course. Agibot reports ~104 ms per continuous action group on an RTX 5090.
- IDM — an inverse dynamics model that converts the imagined future into smooth joint motions. Crucially, this is what makes unlabeled data usable (more below).
- Tier 1: 39,000 h of "instruction–video" pairs for pre-training the world model (includes 3,000 h of embodiment-free first-person and human manipulation video, no action labels).
- Tier 2: 32,000 h of robot trajectories for pre-training the IDM (includes 2,000 h of failed operations and real deployment return-flow data — no success labels required).
- Tier 3: 30,000 h of complete "instruction–video–action" data for joint training.
- Mid-trajectory instruction change. Told to grab a green cup, the arm stops ~1.5 cm short of contact when the command switches to "take the blue cup" — it pauses ~1.5 seconds, abandons the original trajectory, and reorients. A similar test swaps hands mid-reach: right arm retracts, left arm completes the grasp.
- Resisting counterintuitive associations. A cup, a shoe, and a shoebox on the table; instruction: "put the cup in the shoebox." The model ignores the "shoes go in boxes" visual association and follows the literal instruction.
A Hidden Trap: The Validity Gap
Physical manipulation is inherently multi-solution: approaching a cup from the left or the right are both valid, but human demonstrations record only one path. If the visual planner imagines a right-side approach while the training demonstration went left, conventional training forces the robot to watch one imagining while executing the opposite — a supervision conflict that averages strategies into uselessness.
KASO (Knowledge-Aligned Selective Optimization) addresses this: during training, the planner generates multiple candidate futures; the action network first selects the one aligned with the real demonstration, and only that one updates the weights. In real-robot ablations, KASO raised grasp success from 22.5% to 37.5%.
Four Data Scales, Three Lines of Evidence
Same training recipe, only the final-stage data scale varies: 300 / 1,200 / 5,000 / 30,000 hours.
| Training scale | G1-OP success | G2-90D success | Non-zero tasks (G1-OP) | |---|---|---|---| | 300 h | 17.1% | 13.4% | 39 / 100 | | 1,200 h | 22.6% | 21.0% | 51 / 100 | | 5,000 h | 27.3% | 23.5% | 62 / 100 | | 30,000 h | 44.1% | 31.1% | 76 / 100 |
Three findings:
1. Skills light up one by one. Fine manipulations — folding towels, nesting paper cups, capping pens, flower arranging — score zero at small scale and only show non-zero success at 30,000 hours. 2. The curve hasn't saturated. Success keeps climbing from 5,000 to 30,000 hours with no visible plateau. 3. Rich embodiments lift scarce ones. G1-OP supplies over half the training data; the wheeled G2-90D accounts for under 2%, yet still gained 17.7 percentage points. Tasks where G2 has under one hour of data benefited from spatial dynamics learned on other robots.
Agibot also reports a data audit: zero-shot success correlates log-linearly with training hours per skill category. Wiping tables got 824 hours → 76.7% reliability; sweeping with a broom got 64 hours → 3.3%.
Why Not Adapt a Text-to-Video Model?
Most peers start from a pretrained video generation model and bolt on an action head — fast, and it inherits priors from a million hours of internet video.
Agibot's stated objection is a mismatched objective function: video models optimize for visually plausible frames, robots need executable, physically precise actions. The goals overlap early but diverge late — fingers two centimeters off a cup barely change a video's look, but decide grasp success. Pre-training from scratch means building the data pipeline yourself, which motivates the three-tier architecture below.
Three-Tier Data Architecture: Even Failures Count
The IDM is the linchpin: given before/after frames, it recovers what action happened in between. Unlabeled, instruction-free, even failed trajectories yield action supervision. Robots' real-world mistakes don't have to be thrown away.
Two Standouts from Qualitative Testing
Open-Source Rivalry: Unitree on the 10th, Agibot on the 11th
| Dimension | Unitree UnifoLM-WLA-1.0 | Agibot GE-Act 2.0 | |---|---|---| | Release | Sept 10 | Sept 11 | | Scale framing | 6B parameters | Params undisclosed; data scale disclosed | | Coverage | Tabletop + whole-body mobile manipulation | 100 atomic tasks, 20 skill categories | | Public emphasis | Cross-task, cross-end-effector generalization | Data scaling curve and component methods |
The evaluation is deliberately strict: no fine-tuning on test tasks, no demonstrations, direct real-robot zero-shot runs across 100 atomic tasks, 20 skill categories, and two embodiments — with scenes, backgrounds, lighting, and object instances unseen in training.
Industrial Context
Agibot's Spirit G2 is now running routinely on the Longqi (LCFC-partner Longqi Technology) Nanchang factory line — described as the world's first scaled embodied-AI deployment in precision 3C manufacturing — with plans to expand to 100 units in Q3 2026. A Serbia plant with Minth Group began production August 29, targeting 1,000–2,000 humanoids per year.
On benchmarks, Unitree founder Wang Xingxing has defined embodied AI's "ChatGPT moment" as completing 80% of tasks via natural instructions in 80% of unfamiliar scenarios — estimated at 2–3 years at best, 5–10 at worst.
Put together: 44.1% zero-shot is still more than half away from that bar, and production-line return-flow data is exactly the fuel for closing the gap — the source of the 2,000 hours of failure and deployment data in tier two.
What 44.1% Actually Means
In unconstrained environments, 44.1% zero-shot success is far from "ready to work." Agibot itself hedges: the experiments only show that, within their own architecture, data system, and evaluation scope, more training data improves performance on unfamiliar tasks. More embodiments, long-horizon tasks, and unstructured environments remain untested.
Three things to watch:
1. The curve didn't flatten from 5,000 to 30,000 hours — does the next data increment sustain the gains? 2. Can the cross-embodiment transfer (17.7 points on <2% data) replicate to third and fourth embodiments? 3. As these open-source models spread, software moats for smaller robot makers shrink while data moats grow — making upstream data collection and simulation the more certain bets.
Sources
1. Humanoids Daily, "AGIBOT Open-Sources GE-Act 2.0, Challenging Unitree in the Open Embodied AI Race" (2026-09-11): CoAE/SVP/IDM architecture, KASO ablation 22.5%→37.5%, four-scale data and cross-embodiment transfer. 2. Shenzhen Economic Daily / Duchen, "Agibot releases GE-Act 2.0: 100× training data, higher robot zero-shot success" (2026-09-09): 100 tasks, 20 skills, two embodiments, zero-shot no-fine-tune setting. 3. Agibot official release and GE-Act 2.0 technical paper (arXiv:2609.05588): three-tier data architecture, CoAE compression to 24 tokens, 104 ms on RTX 5090. 4. Unitree open-source announcement for UnifoLM-WLA-1.0 (2026-09-10) and related Humanoids Daily coverage: 6B parameters, cross-task/cross-end-effector generalization. 5. Industry daily roundup of the week's world-model releases (2026-09-12): GE-Sim 2.0, UnifoLM-X2-1.0, HERON-World Model, VWA timeline.